Health Insurance and the Demand for Medical Care

Evidence from a Randomized Experiment

Manning, Newhouse, Duan, Keeler & Leibowitz

February 7, 2024

Introduction

  • How does health insurance affect the demand for medical services?

  • Is health insurance to blame for the rapid increase in medical care costs (4% per year) and the rise in the share of GNP devoted to medical care (4.4% in 1950 to 10.7% in 1985)

    • spread of health insurance \(\implies\) increase in demand \(\implies\) increase in price and quantity demanded of medical services
  • Main issue: Insurance is endogenous.

    • Those who expect to demand more medical services have incentives to get more complete insurance

Introduction

  • Contribution: Literature disagreed on the magnitude of the demand response to changes in insurance, price elasticity of demand (or coinsurance elasticity of demand)
    • Elasticity estimates varied from -0.1 to -2.1 at the mean
  • Methodology: The authors use data from a randomized experiment, the Rand Health Insurance Experiment (HIE, 1974-1977), and a
    • Four-Equation Model to exploit the characteristics of the distribution of medical expenditure

Main Findings

  • A catastrophic insurance plan reduces medical services expenditures by 31% relative to zero out-of-pocket price
  • Price elasticity is approximately -0.2
  • Reject the hypothesis that less favorable coverage of outpatient services increases total expenditure (e.g. by decreasing preventing care or inducing hospitalization)

Data

  • The US Federal Government conducted the Rand Health Insurance Experiment (HIE) from 1974 to 1977
  • Enrolled families in six sites selected to reflect varied demographics and medical system complexities
  • Assigned families to different insurance plans with varying cost-sharing structures
  • Utilized the Finite Selection Model to achieve balance across plans while retaining randomization
  • Unit of analysis: Person-year, with control variables for demographics, economics, and health factors

\(\blacktriangleright\) Demographics and Medical System Complexities

Insurance Plan Variability

  • The 14 insurance plans varied in two dimensions of cost sharing:

    1. coinsurance rates1 (Free plan, 25%, 50%, and 95%) and
    2. annual out-of-pocket limits (Maximum Dollar Expenditure MDE of 5%, 10%, or 15% of family income, up to $1,000)
  • Covered expenses included nearly all medical services (excluded services included non-preventive orthodontia and cosmetic surgery)

  • Individual deductible: free inpatient care, 95% coinsurance rate for outpatient services, and $150 expense limit per person ($450 per family)

Sample

  • Random sample of each site’s nonaged population
  • Dependant variable: Medical services (other than outpatient psychotherapy and dental services)
  • Independent variables:
    • Five insurance plan categories: Free plan, 25%, 50%, and 95% coinsurance rate plans, plus the individual deductible plan
    • Controls: age, sex, race, family income, health status, family size, and site

NUMBER OF PERSONS AT ENROLLMENT AND NUMBER OF PERSON-YEARS IN ESTIMATION SAMPLE
Insurance Plan Dayton Seattle Fitchburg Franklin County Charleston Georgetown Enrollment Total (Persons) Estimation Sample Total (Person-Years)
Free 301 431 241 297 264 359 1893 6822
25 Percent 260 253 125 152 146 201 1137 4065
50 Percent 191 0 56 58 26 52 383 1401
95 Percent 280 253 113 162 146 166 1120 3727
Individual Deductible 105 285 188 220 196 282 1276 4175
Total 1137 1222 723 889 778 1060 5809 20190

Empirical Approach

  • Main approach: four-equation model (Duan et al., 1983)
    • vs. simple means (ANOVA)
  • The four-equation model exploits three characteristics of the medical expenses distribution:
    1. large proportion of nonusers during the year,
    2. among users, the distribution is highly skewed,
    3. different distributions for inpatient and outpatient users
  • ANOVA estimates, imprecise but consistent; Four-Eq Model, consistent with lower mean square error

Four-Equation Model

  • Participants grouped into: 1) nonusers, 2) outpatient-only users, and 3) any inpatient service users
  1. Probit for probability of using any medical service (nonusers)
  2. Probit for conditional probability of inpatient stay, given participant is user (inpatient vs. outpatient)
  3. OLS log of total annual medical expenses for outpatient-only users and
  4. OLS log of annual medical expenses for any inpatient service users (skewness)

Expected Medical Expense

\[ E(\text{Medical Expense}_i) = \hat{p}_i \left[ (1 - \hat{\pi})\exp(x_i\hat{\beta}_3)\hat{\phi}_3 + \hat{\pi}\exp(x_i\hat{\beta}_4)\hat{\phi}_4 \right] \]

  • \(\hat{p}_i\): probability of any medical use
  • \(\hat{\pi}_i\): conditional probability for a medical user to have any inpatient use
  • \(\exp(x_i\hat{\beta}_3)\hat{\phi}_3\): conditional expense for medical services if outpatient only
  • \(\exp(x_i\hat{\beta}_4)\hat{\phi}_4\): conditional expense for medical services if any inpatient
  • \(\hat{\phi}_3, \hat{\phi}_4\): “smearing” retransformation of the error terms

\(\blacktriangleright\) Semearing Retransformation \(\blacktriangleright\) Error Term Correction \(\blacktriangleright\) Selection Models

SAMPLE MEANS FOR ANNUAL USE OF MEDICAL SERVICES PER CAPITA

Plan Face-to-Face Visits Outpatient Expenses (1984 $) Admissions Inpatient Dollars (1984 $) Prob. Any Medical (%) Prob. Any Inpatient (%) Total Expenses (1984 $) Adjusted Total Expenses (1984 $)a
Free 4.55
(.168)
340
(10.9)
.128
(.0070)
409
(32.0)
86.8
(.817)
10.3
(.45)
749
(39)
750
(39)
25% 3.33
(.190)
260
(14.70)
.105
(.0090)
373
(43.1)
78.8
(1.38)
8.4
(0.61)
634
(53)
617
(49)
50% 3.03
(.221)
224
(16.8)
.092
(.0116)
450
(139)
77.2
(2.26)
7.2
(0.77)
674
(144)
573
(100)
95% 2.73
(.177)
203
(12.0)
.099
(.0078)
315
(36.7)
67.7
(1.76)
7.9
(0.55)
518
(44.8)
540
(47)
Ind. Deductible 3.02
(.171)
235
(11.9)
.115
(.0076)
373
(41.5)
72.3
(1.54)
9.6
(0.55)
608
(46)
630
(56)
χ2(4)b 68.8 85.3 11.7 4.1 144.7 19.5 15.9 17.0
p-value χ2 <.0001 <.0001 .02 n.s. <.0001 .0006 .003 .002

Results (Simple Means)

  • Per capita expenses on the free plan (no out-of-pocket costs) are 45% higher than the 95% coinsurance rate.
    • The higher the cost share, the lower the per capita expenses.
  • Outpatient expenses on the free plan are 67% higher than the 95% plan, outpatient visit rates are 66% higher.
    • Cost sharing affects the number of medical contacts, rather than the intensity.

Results (Simple Means)

  • Largest significant drop in outpatient between the free plan and 25% percent plan

  • Inpatient services use does not change significantly between the 25, 50, and 95% plans

Results (Simple Means)

Individual plan (free inpatient care, 95% outpatient coinsurance, $150 Max per person) participants behave differently

  • Expenditure is significantly lower than free plan;
    • the overall response is: one-third reduction in outpatient expenses
    • less than one-tenth reduction in inpatient expenses
  • Inpatient service usage lies between the free plan and coinsurance plans, suggesting non-trivial cross-price elasticity between inpatient and outpatient services

VARIOUS MEASURES OF PREDICTED MEAN ANNUAL USE OF MEDICAL SERVICES, BY PLAN

Plan Likelihood of Any Use
(%)
One or More Admissions
(%)
Medical Expenses
(1984 $)
Free 86.7
(0.67)
10.37
(0.420)
777
(32.8)
Family Pay
25 Percent 78.8
(0.99)
8.83
(0.379)
630
(29.0)
50 Percent 74.3
(1.86)
8.31
(0.400)
583
(32.6)
95 Percent 68.0
(1.48)
7.75
(0.354)
534
(27.4)
Individual Deductible 72.6
(1.14)
9.52
(0.529)
623
(34.6)
Note: Standard errors are shown in parentheses.

Results (Four-Eq. Model)

  • Mean predicted expenditure is 46% higher in the free care than in the 95% plan

  • Largest response to plan occurs between free care and 25% plan

  • 87% of participants in the free plan are predicted to use any medical services during the year, while only 68% of participants in the 95% plan are

    • Same as in simple means, adding covariates does not change estimated probabilities of medical service use
  • Outpatient-only cost sharing (individual deductible plan) reduces total expenditures relative to free care by reducing the likelihood of any use

\(\blacktriangleright\) Subgroups \(\blacktriangleright\) Other Subgroups \(\blacktriangleright\) Other Results

Literature Estimates of Demand

Three different methods to estimate price elasticity comparable to the literature:

  1. Episodes. Changes in demand for episodes of treatment when individuals are more than $400 from their limit

  2. Indirect utility function and applying it to total expenditure

  3. Average coinsurance rates

  • The three methods suggest price elasticities for a constant coinsurance rate between -0.1 to -0.2. Values are similar to the lower range of the nonexperimental literature.

Discussion

  • Does health insurance explain the sustained rise in medical expenditure?
    • Demand response estimates can only account for about a tenth of the factor of 7 change in health expenditure in the post-World War II period (1950-1984).
    • Changes in income can neither account for the increase
    • Technological change? host of new medical products, like dialysis or kidney transplantation
  • Welfare loss from moving from a universal 95 percent plan to the free care plan is $37 to $60 billion; expenditure on medical services in 1984 by under 65 population was $200 billion

Discussion

  • Traditionally inpatient services coverage is more generous relative to outpatient services coverage

  • Lack of insurance for outpatient services might deter individuals from seeking care (reducing medical services expenditure) when their illness can be treated relatively cheaply

  • More generous coverage for inpatient services might lead physicians to hospitalize patients who could be treated on an outpatient basis increasing social expenditure

  • Are outpatient and inpatient services substitutes or complements?

Discussion

  • The authors reject the hypothesis that increased coverage of outpatient services, holding constant the coverage of inpatient services, will reduce expenditure (more coverage, less expenditure)

  • Mean expenditure on individual plan (free inpatient, costly outpatient) is 20 percent less than the mean on the free care plan (free inpatient, free outpatient) (less coverage, less expenditure)

  • Moreover, outpatient deductible (less coverage) also decreases hospital admissions (inpatient care), suggesting inpatient and outpatient services are complements

Appendix

Selection of Six Cities

  • Six cities chosen to represent the four census regions.
    • Dayton, Ohio
    • Seattle, Washington
    • Fitchburg, Massachusetts
    • Franklin County, Massachusetts
    • Charleston, South Carolina
    • Georgetown County, South Carolina
  • To represent
    • City Sizes to capture medical delivery system complexities
    • Waiting times to appointment and physician per capita ratios to test sensitivity of demand elasticities to nonprice rationing
    • Geographical Diversity: urban and rural sites in the North and the South

\(\blacktriangleright\) Back

Smearing Retransformation

Can’t use normal theory restransformation from logs to dollars \((\exp(\sigma^2/2))\) because the error distributions of the OLS log regressions deviate from the normal assumption.

Nonparametric estimate of the retransformation of factors (Duan, 1983),

\[ \hat{\phi_j}=\Sigma_i \exp(\hat{\varepsilon_{ij}})/n_j, \; j=3,4, \]

where \(n_j\) is sample size of equation \(j\), \(\hat{\varepsilon}_{ij}=\ln(y_{ij})-x_{ij}\hat{\beta}_j\), and \(\hat{\beta_j}\) is the OLS estimate, \(i\) indexes the person.

\(\blacktriangleright\) Back

Nonparametric Correction in the Error Terms

  • Results show a mean value for each plan
    • mean within plans over predicted values obtained by alternatively placing participants on each plan
  • Clustered Standard Errors by family and over time

\(\blacktriangleright\) Back

Comparison with Selection Models

  • Types:
    1. Tobit model.
    2. Adjusted Tobit model.
    3. Sample selection models.
  • Characteristics:
    • Model correlation between probability of any use and level of use explicitly.
    • the four-equation model is not nested within selection models.
  • Comparison:
    • Four-equation model demonstrated less bias and statistical indistinguishability in empirical validation.

\(\blacktriangleright\) Back

Use by Subgroups

  • Income:
    • The probability of any use of medical services increases with income within each plan, but these estimates are influenced by the income-related upper limit in out-of-pocket expenses
    • A cleaner test —the individual deductible plan because its deductible is not related to income— shows the differences in medical expenses due to income are not significant
  • Age groups:
    • Similar outpatient response to insurance plans for children (age ≤ 18) and adults.
    • Children less plan-responsive for inpatient care.

Other Subgroups

  • Health Status:
    • No differential response to health insurance coverage observed between healthy and sickly individuals.
    • Lack of interaction between plan and health status suggests sickly individuals exhibit more discretion at the margin.
  • Sites:
    • No differences in response to insurance coverage among sites.
  • Period of Enrollment:
    • Duration of enrollment does not significantly impact spending or plan responsiveness.

Other Results

  • Dental Results:
    • Dental services show greater responsiveness to plan in the first year.
    • Similar magnitude of responsiveness as other medical services in subsequent years.
  • Health Maintenance Organization (HMO) Results: A randomized group of participants was assigned into an HMO, the Group Health Cooperative of Puget Soung in Seattle. The HMO Experimentals. This group was given a plan of benefits identical to the free-for-service (FFS) plan.
    • No evidence of selection observed in the HMO studied.
    • HMO Experimentals exhibit lower inpatient use and lower expenditures compared to free fee-for-service plan.
    • Health status outcomes comparable between HMO and free fee-for-service plan.